Deep Visual SLAM - how deep learning improves SLAM
Visual Simultaneous Localisation And Mapping (SLAM) has been one of the most relevant research topics in robotics for the last two decades. Many algorithms have been developed to make these methods more reliable, computational efficient and accurate. These solutions can be distinguished in traditional methods and deep learning-based methods. Traditional methods mainly use handcrafted algorithms, while deep learning methods use CNN and transformer-based architecture. Despite the very promising results and their easy integration with other tasks (such as segmentation), learning methods have not replaced traditional algorithms mainly because of their technical requirements, limiting their widespread deployment.
This MSc project aims to critically assess the performance of these methods and design a solution leveraging both methods and outperforming their current limitations.
The objectives of this MSc project can be summarised as follows:
1) Comparison and analysis of existing methods (both traditional and deep learning ones) using some meaningful real datasets.
2) Analysis of the strengths and weaknesses of the methods
3) Design of a new solution leveraging strengths of existing methods: for the designed solution, a hybrid approach using both traditional and deep learning methods is foreseen
4) Testing of the solution in different operational conditions.